"AI restores your photo" sounds like magic, and most sites are happy to leave it there. We'd rather you understand what's actually happening — because knowing how it works also tells you exactly what it can't do.
Every restoration model is a neural network trained the same basic way: take millions of clean photographs, artificially degrade copies of them (blur, shrink, add grain, compress, fade), and train the network to turn each degraded copy back into its clean original. Repeat at enormous scale, and the network learns two things: what degradation does to real photos, and what the sharp versions of real-world textures — skin, hair, fabric, brick, foliage — actually look like.
When you then hand it your faded scan, it applies that learned knowledge: "this smudge, in this context, is what a sharp eye usually looks like after fifty years of fading — here is that eye, rebuilt." It doesn't retrieve your photo's lost pixels from anywhere; it synthesises new pixels that are statistically consistent with what remains. That distinction matters, and we'll come back to it.
The sharpen & upscale mode uses a super-resolution model of the Real-ESRGAN family — a widely used, well-studied open architecture. Given an image, it outputs a version with 4× the pixels in each direction, filling in the new pixels with learned texture rather than the smooth interpolation a normal photo editor uses.
The difference is easy to see. Classic enlarging (bicubic interpolation) averages neighbouring pixels, so a 4× enlargement is just the same blur, bigger. A super-resolution model instead asks, for every region, "what sharp structure most plausibly produced these blurry pixels?" — and draws that. Edges become genuinely crisp, film grain and JPG blockiness get cleaned up in the process, and textures come back with believable detail. Real-ESRGAN in particular was trained on aggressively degraded images (blur + noise + compression stacked together), which is why it copes so well with old scans, whose damage is always a messy combination rather than one clean defect.
Two practical notes from how we run it: large photos are processed in overlapping tiles so the GPU can handle them, and the input is capped at a sane size — feeding a 60-megapixel scan through a 4× model would otherwise produce absurd files. The better your input, the better the output, which is why a good scan matters — see our scanning guide.
The colourise mode does something subtly different. A black & white photo already contains the brightness of every pixel — what's missing is the colour information. The model keeps the brightness channel untouched and predicts only the two colour channels, using context: it has learned from millions of colour photos that skies are blue, foliage is green, skin falls in a certain range, wood is brown. Because the original brightness is preserved, the result keeps every bit of the photo's real structure — the AI only "paints within the lines" it's given.
Colourisation is a plausible prediction, not a historical record. The model cannot know that your grandmother's dress was actually red — it will pick a colour that's statistically likely for a dress of that shade in that era of photo. Objects without strong colour conventions (cars, clothing, painted walls) are exactly where it guesses most. Treat a colourised photo as an interpretation that makes the past feel closer, not as evidence of what colours things were.
Faces get their own mode (restore faces) for a good reason: humans are hyper-sensitive to faces. A generic upscaler that does a beautiful job on a brick wall can leave a face subtly "off" in a way everyone notices instantly. Face-restoration models (ours is from the CodeFormer family) are trained exclusively on faces, so they carry much stronger prior knowledge about facial structure — how eyes, teeth, and skin texture look at every age and angle. The pipeline detects each face in the photo, restores it with the specialist model, and blends it back into the surrounding image.
These models have a "fidelity" dial: lean one way and the output stays maximally faithful to the degraded input (but restores less); lean the other way and faces come out beautifully sharp (but the model invents more). We run a deliberately faithful setting, because for family photos, resemblance beats beauty.
On a mildly blurred face, the reconstruction is guided by plenty of surviving pixels and stays very true. On a heavily degraded face — a tiny smudge of a face in a group photo — the model has to invent most of the detail, and the result can drift from what the person really looked like. It will look like a real face; whether it's exactly their face is not guaranteed. If identity matters, compare against other photos of the same person.
All of the above runs on our own GPU hardware — which is why the tool is free, with no account and no watermark, and why your photo can be deleted from the machine right after processing. Upload a photo, drag the before/after slider, and judge the result with exactly the expectations this page gave you. That's the deal we prefer: impressive results, honestly framed.